Fast recognition of instantaneous states of pool boiling based on deep transfer learning
The boiling process is an efficient heat transfer method that creates a boiling crisis.Instantaneous recognition and prediction of the boiling state are necessary to maintain efficient heat dissipation and prolong the life cycle of high-power-density electronic devices.Benefiting from the rapid advancements in artificial intelligence,intelligent monitoring and rapid recognition technologies have been gradually applied to develop efficient boiling-state detectors.Traditional machine learning methods,such as pure data-driven methods,encounter difficulties in avoiding their excessive dependence on data.However,in practice,models frequently experience generalization bottlenecks due to changes in working conditions and data scarcity.To resolve these issues,an instantaneous boiling-state recognition method based on deep transfer learning is proposed in this paper.First,a state recognizer OrigCNN is constructed based on the convolutional neural network and subsequently trained using Dataset A,obtained from our pool boiling experiments.The accuracy of the test is up to 100％.Considering the generalization bottleneck of the source model OrigCNN,the transfer learning technology is applied to further improve OrigCNN by constructing a deep transfer model TLCNN,with"Dataset A"as the source domain and the publicly available"Dataset B"and"Dataset C"as the target domain.10％,5％,2.5％,and 1％of Dataset B and Dataset C are used to construct small sample datasets for transfer training.The test results show that the amount of sample data for transfer learning positively correlates with the prediction accuracy of TLCNN.The TLCNN test accuracy reached 99.83％when 5％Dataset B(132 photos)was used for transfer training,and the false negative rate of critical state detection was 0.38％,demonstrating the effectiveness and reliability of the TLCNN models in actual scene switching.Furthermore,the deep transfer learning method TLCNN proposed in this research exhibits high identification efficiency on the millisecond scale using a single computer device,which is of considerable importance for developing real-time boiling transient state recognizer and digital twin software tools.
NETL
NSTL
万方数据
